The 3D digital twin of telecommunication infrastructures, including the 3D model of metallic towers, ground technical rooms, and associated equipment (e.g.; Low Voltage Distribution Panel, Base Transceiver Station, or cells), is an essential tool used during the engineering and maintenance phases. Indeed, this 3D digital twin allows to simulate the location of future equipment, and needs to be kept up to date in order to avoid decision making based on an obsolete representation.
Thus, the thesis will aim at detecting discrepancies between the real telecommunication infrastructure and its 3D digital twin using common vision sensors such as those embedded in smartphones or in civil UAVs. From these captures made from a number of viewpoints, the implemented solution will have to identify the different telecommunication equipment’s positioned on the metallic pylon or in the technical room, in order to match them with those of the 3D digital twin, and indicate the discrepancies between the real infrastructure and its digital twin.
This thesis will mainly rely on machine learning approaches, whether for dense 3D reconstruction, 2D and 3D panoptic segmentation, relocalization, or any other computer vision processing.
This thesis is part of an industrial context and will try to answer a concrete industrial need. Its results will have to be tested in a representative environment.
Expected results:
• A state of the art on discrepancy detection between a 3D model and the real environment from RGB images.
• The publication and implementation of innovative representations and techniques to detect discrepancies between a 3D model and the real environment from RGB images.
• One or more experiments of the solution applied on telecommunication infrastructures.
This CIFRE thesis will be supported by TDF (Vi OLIVET, Digital Lab leader), directed by Pr. Panagiotis PAPADAKIS from IMT Atlantique, and will be part of a project supported by the IRT b<>com (Dr. Jérôme ROYAN, senior scientist and principal architect).
Profile :
The candidate must have excellent computer development skills (C++, Python) and must also demonstrate skills in computer vision (Localization, 3D reconstruction, 2D and 3D segmentation, etc.), and in machine learning (CNN, GaN, NeRF, etc.).